New switch upgrades Cisco high-end server-to-storage computing

Cisco is introducing a gateway switch that boosts the performance of applications that rely on interactions between InfiniBand-clustered servers and network-attached storage- and Fibre Channel-based storage devices.

Cisco Server Fabric Switch (SFS) 3504 can aggregate server-cluster traffic on 20Gbps InfiniBand links and distribute it to storage devices on Ethernet and Fibre Channel networks.

This is an upgrade from the 10Gbps InfiniBand links supported by Cisco’s earlier SFS 3012.

Rather than installing a Fibre Channel host channel adapter in each server in the cluster so it can connect directly to the Fibre Channel storage-area network, each server connects via InfiniBand to the SFS 3504. The SFS 3504 switches this aggregated traffic to the SAN via 4Gbps Fibre Channel ports. This is an upgrade from 2Gbps Fibre Channel ports on the SFS 3012.

The SFS 3504 also takes in Ethernet traffic running over InfiniBand from the same server clusters and switches it to NAS devices via 1Gbps Ethernet ports.

InfiniBand lets machines connect and share data with sub-microsecond latency. In high-performance computing environments, these servers need to connect with storage networks based on both Fibre Channel and Ethernet.

These environments are used commercially by oil and gas companies to choose drilling sites based on crunching seismic data and by car makers to simulate crashes of new models using fewer prototypes.

These Cisco gateway switches compete against gear from Voltaire, QLogic and Xsigo.

The gateway can be outfitted with four-port, 4Gbps Fibre Channel cards or six-port 1Gbps Ethernet cards. Next year Cisco plans to release a two-port 10Gbps Ethernet card for the device.

Fully loaded, the SFS 3504 costs $75,000.

Cisco is also releasing a free software tool that troubleshoots copper cabling in 20Gbps InfiniBand networks. The tool can analyze signal integrity of InfiniBand on these connections prior to running a computing job that involves multiple hosts. If any of the nodes is having a significant problem, it can be fixed or another node can be chosen so the problem doesn’t slow down the whole run, Cisco says.